How did I run a TensorFlow training on the PASCAL VOC dataset in Ubuntu 18.04?

dmitrii_fediuk · May 12, 2019, 9:16pm

Step 1

I have downloaded the PASCAL VOC dataset (the VOCtrainval_11-May-2012.tar file).
Then I extracted the archive’s contents.

Step 2

Then I have generated TFRecord files:

github.com

dmitry-ai/tf-tutorial-pascal-voc/blob/0.0.3/create_pascal_tf_records.sh#L3-L17


input='/w/data/VOCdevkit'
objectDetection='/w/tools/tensorflow/models/research/object_detection/'
output='/w/tutorials/tensorflow/pascal-voc/data/'
declare -A targets=( \
	[train]='pascal_train.record' \
	[val]='pascal_val.record' \
)
for type in "${!targets[@]}"; do
	python ${objectDetection}dataset_tools/create_pascal_tf_record.py \
		--label_map_path=${objectDetection}data/pascal_label_map.pbtxt \
		--data_dir=$input \
		--year=VOC2012 \
		--set=$type \
		--output_path=${output}pascal_${type}.record
done

github.com/tensorflow/models/blob/v1.13.0/research/object_detection/g3doc/preparing_inputs.md#generating-the-pascal-voc-tfrecord-files

Step 3

github.com

dmitry-ai/tf-tutorial-pascal-voc/blob/0.0.4/faster_rcnn_resnet101_voc07.pbtxt#L6-L140


model {
	faster_rcnn {
		# 2019-05-13 What does mean «faster_rcnn» → «num_classes» in TensorFlow? https://dmitry.ai/t/62
		num_classes: 20
		image_resizer {
			# 2019-05-13
			# What does mean «faster_rcnn» → «image_resizer» → «keep_aspect_ratio_resizer» in TensorFlow?
			# https://dmitry.ai/t/61
			keep_aspect_ratio_resizer {
				max_dimension: 1024
				min_dimension: 600
			}
		}
		# 2019-05-13 https://en.wikipedia.org/wiki/Feature_extraction
		# 2019-05-16 What is «feature extraction» in machine learning? https://dmitry.ai/t/73
		feature_extractor {
			type: 'faster_rcnn_resnet101'
			# 2019-05-16
			# What does mean «model» → «faster_rcnn» → «feature_extractor» →
			# «first_stage_features_stride» in TensorFlow? https://dmitry.ai/t/92
			first_stage_features_stride: 16
		}
		first_stage_anchor_generator {
			grid_anchor_generator {
				scales: [0.25, 0.5, 1.0, 2.0]
				aspect_ratios: [0.5, 1.0, 2.0]
				height_stride: 16
				width_stride: 16
			}
		}
		# 2019-05-16
		# What does mean «model» → «faster_rcnn» → «first_stage_box_predictor_conv_hyperparams» in TensorFlow?
		# https://dmitry.ai/t/97
		first_stage_box_predictor_conv_hyperparams {
			op: CONV
			regularizer {
				l2_regularizer {weight: 0.0}
			}
			initializer {
				truncated_normal_initializer {
					stddev: 0.01
				}
			}
		}
		first_stage_nms_score_threshold: 0.0
		first_stage_nms_iou_threshold: 0.7
		first_stage_max_proposals: 300
		first_stage_localization_loss_weight: 2.0
		first_stage_objectness_loss_weight: 1.0
		initial_crop_size: 14
		maxpool_kernel_size: 2
		maxpool_stride: 2
		second_stage_box_predictor {
			mask_rcnn_box_predictor {
				use_dropout: false
				dropout_keep_probability: 1.0
				fc_hyperparams {
					op: FC
					regularizer {
						l2_regularizer {
							weight: 0.0
						}
					}
					initializer {
						variance_scaling_initializer {
							factor: 1.0
							uniform: true
							mode: FAN_AVG
						}
					}
				}
	  		}
		}
		second_stage_post_processing {
			batch_non_max_suppression {
				score_threshold: 0.0
				iou_threshold: 0.6
				max_detections_per_class: 100
				max_total_detections: 300
			}
			score_converter: SOFTMAX
		}
		second_stage_localization_loss_weight: 2.0
		second_stage_classification_loss_weight: 1.0
	}
}
train_config: {
	# 2019-05-17
	# 1) What does mean «train_config» → «batch_size» in TensorFlow?  https://dmitry.ai/t/98
	# 2) What is the optimal batch size for a TensorFlow training? https://dmitry.ai/t/100
	# 3) How to estimate the maximum allowed batch size for a TensorFlow training? https://dmitry.ai/t/102
	batch_size: 1
	data_augmentation_options {random_horizontal_flip {}}
	gradient_clipping_by_norm: 10.0
	fine_tune_checkpoint: "/w/data/pre-trained/faster_rcnn_resnet101_coco_2018_01_28/model.ckpt"
	from_detection_checkpoint: true
	optimizer {
		momentum_optimizer: {
			learning_rate: {
				manual_step_learning_rate {
					initial_learning_rate: 0.0001
					schedule {
						step: 500000
						learning_rate: .00001
					}
					schedule {
						step: 700000
						learning_rate: .000001
					}
				}
			}
			momentum_optimizer_value: 0.9
		}
		use_moving_average: false
	}
	# 2019-05-17 What does mean «train_config» → «num_steps» in TensorFlow?  https://dmitry.ai/t/99
	num_steps: 10000
}
train_input_reader: {
	tf_record_input_reader {
		input_path: "/w/tutorials/tensorflow/pascal-voc/data/pascal_train.record"
	}
	label_map_path: "/w/tools/tensorflow/models/research/object_detection/data/pascal_label_map.pbtxt"
}
eval_config: {
	num_examples: 4952
}
eval_input_reader: {
	tf_record_input_reader {
		input_path: "/w/tutorials/tensorflow/pascal-voc/data/pascal_val.record"
	}
	label_map_path: "/w/tools/tensorflow/models/research/object_detection/data/pascal_label_map.pbtxt"
	shuffle: false
	num_readers: 1
}

Step 4

github.com

dmitry-ai/tf-tutorial-pascal-voc/blob/0.0.4/training.sh#L2-L8


objectDetection='/w/tools/tensorflow/models/research/object_detection/'
python ${objectDetection}model_main.py \
	--pipeline_config_path=faster_rcnn_resnet101_voc07.pbtxt \
	--model_dir=../run/2019-05-17 \
	--num_train_steps=10000 \
	--sample_1_of_n_eval_examples=1 \
	--alsologtostderr

Step 5

tensorboard --logdir=/w/tutorials/tensorflow/pascal-voc/run/2019-05-17

dmitrii_fediuk · May 12, 2019, 11:34pm

See also: How did I run a TensorFlow training on the PASCAL VOC dataset in Windows 7?